Last month I spent a few days going deep on a space I knew nothing about: using genetic and evolutionary models to orchestrate thousands of scientific experiments that learn from each other. Labs running AI-directed research loops, generating hypotheses, running trials, feeding results back into the model. A year ago, that kind of domain survey took weeks — papers, experts, conferences. This time: three days, a working mental model, and a research note with sources and second-order questions. I'll write about that space separately. The point here is what made it possible.
The difference wasn't speed. It was cognitive load.
High-leverage knowledge work — investing, operating, research, strategy — has one real bottleneck: cognitive load. Every context switch burns working memory. Every research tangent fragments attention. Assembly tasks consume cycles that should go to synthesis. The constraint isn't time. It's how long you can stay in judgment mode before you have to drop back into gathering mode.
I built a system to answer that. The insight it forced: agentic principles aren't just for engineers. They're a meta-skill for anyone whose product is thinking.
The Zed Frame — and What It Misses
Zed's "agentic engineering" framing nailed something real. Four principles define it:
- Quality responsibility stays with the engineer
- Craftsmanship remains critical. AI amplifies taste, doesn't replace it.
- Directing agents is a craft (thinking in systems, not just writing prompts)
- Rapid feedback loops between human and agent
This is right for developers. But it undersells the point. The same principles apply to any domain where the work is high-cognitive-load and the output is judgment: investors synthesizing thousands of signals into high-conviction decisions, operators coordinating across teams, researchers building mental models from fragmented literature.
Engineering didn't invent problem-solving. It perfected a version of it. The deeper abstraction: any field where thinking is the product can be restructured around these principles. That's what I mean by agentic X — fill in your discipline.
How It Works: The Soba System
I built Soba around one principle: reduce friction through system design, not willpower.
Three stages:
Stream capture. One place to log everything. Meeting notes, ideas, tasks, observations, whatever's passing through. Daily note, stream section, append-only. I dump my stream of thoughts in the moment, wherever I am. The goal is zero capture friction: don't lose the thought, don't curate it yet.
Daily closeout. End-of-day triage. Stream items become tasks, research seeds, or synthesized into day essence. An agent-driven review within the context of my priorities and open tasks. This is the most powerful part: it keeps you on track without requiring you to manually track things. The agent holds the context; I make the calls.
Morning briefing. Context prep for the day. Open items, calendar, research deltas filtered through my priority register and knowledge base. The day starts with a full picture, not a cold start.
The system is built around me, not to replace me. I own quality. The agent handles context assembly and cross-referencing. I handle judgment. We get better over time: execution, feedback, tighter system, do more, better.
What This Looks Like: Idea to Education in 24 Hours
Before Soba, exploring a new space meant weeks of tab-hopping. Read a paper, meet an expert, chase citations, build mental model, repeat. Time to basic fluency: 1–2 weeks.
With Soba:
- Capture idea in daily stream: "Explore security implications of terminal access in agents"
- During closeout, convert to research task — add context if I have it
- Morning briefing: Soba runs thesis expansion, sources, counterarguments, related work
- I review synthesis, challenge claims, ask second-order questions, share with others for perspective
- Soba updates draft with responses
24 hours from capture to educated, with sources and a working mental model.
A big one for me personally: the cold-start problem. Starting a new domain no longer means blank page and tab soup. There's always a scaffold.
What This Actually Unlocks
Agentic engineering unlocks latent productivity. Agentic operating unlocks something harder to measure: the ability to go wider and deeper at the same time, without tradeoff.
Normally, depth and breadth compete. You go deep on one thread and lose the others. With Soba:
- Research runs in parallel while I stay in synthesis mode
- Draft scaffolding emerges from conversation, no blank page
- I focus on second-order questions instead of first-order assembly
Better essays. Better decisions. Better thinking. The quality goes up because the friction goes down.
The Obvious Objection
Why wouldn't this stay niche — just a personal productivity system for people technical enough to build one?
Two reasons.
First, the same thing was said about Zed's agentic engineering frame. "This is for developers who already live in the terminal." That turned out to be underselling it — the principles work because they're structural, not tool-specific. Quality accountability, craftsmanship, system-level thinking, tight feedback loops: these apply whenever the output is judgment and the bottleneck is cognitive load.
Second, the tools are getting simpler fast. You don't need to be a developer to run Soba. You need to care enough to design a system. That bar is high in a different way, but it's not technical. The people who build personal operating systems early get compounding advantage — better context, better decisions, better institutional memory. That gap widens over time.
The question isn't whether this stays niche. It's who builds the platform that makes it accessible.
The Surface This Exposes
Building Soba led to an unexpected finding.
When you give an agent terminal access to run research, schedule tasks, and update your knowledge graph, you create a new attack surface. Prompt injection becomes privilege escalation. A malicious webpage can hijack your agent's context and run commands on your machine. Not hypothetical. Structural.
This isn't a Soba problem. It's a property of any agentic system with tool access. The terminal is the new browser. I'll write about this separately.